🤖 Deep Research Bot
Performs an in-depth online research to answer user queries. Split questions into sub-queries for deeper scans.
Created by HARPA AI
Updated on Nov 11, 08:24
Installed 217 times
How to Use
- Get HARPA AI from https://get.harpa.ai
- Open HARPA AI with Alt+A on this page
- Click IMPORT COMMAND
Content
- type: say
message: >-
**Tip:** This command performs a thorough in-depth online search for a given
query or topic and is best used for research. It takes time and is token
hungry (~20k tokens on GPT-4o-mini per search depth).
- param: userQuery
message: ' Type your search query:'
type: ask
options: null
default: ''
- param: depth
message: >-
**Set the search depth.**
Search depth indicates how many search pages will be analyzed. Each depth
level generates an extra sub-query, improving search quality.
options:
- label: 🔍 SEARCH
value: 1
- label: 2 SUB-QUERIES
value: 2
- label: 3 SUB-QUERIES
value: 3
- label: 4 SUB-QUERIES
value: 4
type: ask
default: ''
optionsInvalid: false
- steps:
- prompt: >-
You are a Research Agent. I need you to break down my search query into
{{depth}} clarifying queries in {{language}}, the answers to which are
necessary to form a detailed response to my query: "{{userQuery}}"
Follow the instructions:
- The first sub-query should be as close as possible to My Query in
wording. If my query is short and simple - repeat it in the first
sub-query.
- Subsequent sub-queries should clarify the first one.
- Each sub-query should include keywords from [my search query]. For
example, if asking about a city, include it in all sub-queries to find
relevant info.
- Sub-queries should be brief. Imagine typing a brief Google query for a
precise answer.
- Do not repeat my command and instructions.
- Response in JSON array with {{depth}} objects (queries).
- Write nothing other than the JSON.
Example input: "What is the history of the steam engine, its industrial
impact, and who invented the first practical version and when?"
Example JSON response:
[
{
"query": "Steam engine history and industrial impact"
},
{
"query": "First practical steam engine inventor and date"
}
]
JSON response:
param: subqueries
condition: '{{depth}} > 1'
type: gpt
silent: true
- type: calc
func: extract-json
to: subqueries
param: subqueries
index: all
- message: >-
🧩 To give a comprehensive response, your question was split into
**{{subqueries.length}}** sub-queries.
type: say
- steps:
- value: '{{item.query}}'
type: calc
func: set
param: query
format: text
- message: '⏳ Scanning information for sub-query: "**{{query}}**".'
type: say
- param: information
value: '{{serp {{query}}}}'
type: calc
func: set
format: ''
- func: serp.extract-links
to: links
type: calc
from: information
- param: total
value: '{{links.length}}'
type: calc
func: set
format: number
- steps:
- value: '{{page {{item.url}}}}'
type: calc
func: set
param: content
format: ''
- param: pageUrl
type: calc
func: set
format: text
value: '{{item.url}}'
- code: |-
const regex = /^(?:https?:\/\/)?(?:www\.)?([^\/]+)/;
const testString = pageUrl;
const matches = testString.match(regex);
if (matches) {
const hostname = matches[1];
return hostname;
}
param: hostname
type: js
args: pageUrl
timeout: 15000
silent: true
- type: calc
func: increment
param: index
delta: 1
- args: index, total
code: |-
function calculatePercentage(index, total) {
index = Number(index);
total = Number(total);
if (isNaN(index) || isNaN(total)) {
return "Error";
}
return (index / total * 100).toFixed(1) + "%";
}
let result = calculatePercentage(index, total);
return result;
param: percentage
type: js
timeout: 15000
silent: true
- steps:
- prompt: >-
Act as a professional researcher.
Analyze the given [web page content] and summarize useful
information to answer my query: "{{query}}"
Instructions:
- Imagine you are a Google Search engine, collecting
relevant information from various websites to form an answer
to my query in {{language}}.
- Gather only information that will be useful in answering
my question.
- Extract as much useful information as possible, be concise
and don't write more than needed to answer the question.
- Avoid general phrases.
- When referring to someone's opinion about something,
mention the author. Quote if appropriate.
- If you haven't found anything useful, don't make up
information or guess - simply say "No relevant information
found" in {{language}}.
- Do not echo my prompt in your response.
- Respond with a JSON object containing a single field:
"info".
- Write nothing other than the JSON.
Example JSON response:
{
"info": "Write here textual information that would help answer my query. Write as much as you think is necessary, focusing on providing useful information for further processing."
}
[Web page content]: {{content}}.
JSON response:
param: data
type: gpt
isolated: true
silent: true
- index: first
type: calc
func: extract-json
to: data
param: data
- param: data.url
type: calc
func: set
format: auto
value: '{{item.url}}'
- param: data.title
type: calc
func: set
format: auto
value: '{{item.title}}'
- list: array
type: calc
func: list-add
index: last
item: data
- to: SAY STATUS
type: jump
condition: '{{content}} ~= /(?s)^.{200,}$/gm'
label: FETCHED
type: group
- steps:
- type: navigate
url: '{{item.url}}'
waitForIdle: false
silent: false
- type: wait
for: 0.5s
silent: true
- prompt: >-
Act as a professional researcher.
Analyze the given [web page content] and summarize useful
information to answer my query: "{{query}}"
Instructions:
- Imagine you are a Google Search engine, collecting
relevant information from various websites to form an answer
to my query in {{language}}.
- Gather only information that will be useful in answering
my question.
- Extract as much useful information as possible from the
page content.
- Avoid general phrases.
- When referring to someone's opinion about something,
mention the author. Quote if appropriate.
- If you haven't found anything useful, don't make up
information or guess - simply say "No relevant information
found" in {{language}}.
- Do not echo my prompt in your response.
- Respond with a JSON object containing a single field:
"info".
- Write nothing other than the JSON.
Example JSON response:
{
"info": "Write here textual information that would help answer my query. Write as much as you think is necessary, focusing on providing useful information for further processing."
}
[Web page content]: {{page {{query}} limit=30%}}.
JSON response:
type: gpt
isolated: true
param: data
silent: true
- type: calc
func: extract-json
to: data
param: data
index: first
- type: calc
func: set
param: data.url
format: auto
value: '{{item.url}}'
- type: calc
func: set
param: data.title
format: auto
value: '{{item.title}}'
- type: calc
func: list-add
index: last
list: array
item: data
label: NOT FETCHED
type: group
- message: >-
🔍 Analyzed **{{index}} / {{total}}**,
[{{hostname}}]({{pageUrl}})
label: ⛔️SAY STATUS
condition: '{{index}} = {{total1}}'
type: say
- message: 🔍 Analyzed **{{percentage}}**, [{{hostname}}]({{pageUrl}})
condition: '{{index}} != {{total}}'
type: say
label: SAY STATUS
- message: |-
✅ **{{percentage}}** pages scanned for "**{{query}}**".
Last checked page: [{{hostname}}]({{pageUrl}})
condition: '{{index}} = {{total}}'
type: say
label: SAY STATUS
type: loop
list: links
type: loop
list: subqueries
- message: >-
🤖 Analyzed **{{array.length}}** pages from **{{subqueries.length}}**
Google searches.
Now preparing a comprehensive answer to your question.
type: say
label: DEPTH > 1
type: group
condition: '{{depth}} > 1'
- steps:
- message: ⏳ Give me a minute to search through the web for you...
type: say
- value: '{{userQuery}}'
type: calc
func: set
param: query
format: text
- type: calc
func: set
param: information
format: ''
value: '{{serp {{query}}}}'
- type: calc
func: serp.extract-links
from: information
to: links
- type: calc
func: set
param: total
format: number
value: '{{links.length}}'
- steps:
- label: TRYING TO FETCH
type: calc
func: set
param: content
format: ''
value: '{{page {{item.url}}}}'
- type: calc
func: set
param: pageUrl
format: text
value: '{{item.url}}'
- type: js
args: pageUrl
code: |-
const regex = /^(?:https?:\/\/)?(?:www\.)?([^\/]+)/;
const testString = pageUrl;
const matches = testString.match(regex);
if (matches) {
const hostname = matches[1];
return hostname;
}
param: hostname
timeout: 15000
silent: true
- type: calc
func: increment
param: index
delta: 1
- type: js
args: index, total
code: |-
function calculatePercentage(index, total) {
index = Number(index);
total = Number(total);
if (isNaN(index) || isNaN(total)) {
return "Error";
}
return (index / total * 100).toFixed(1) + "%";
}
let result = calculatePercentage(index, total);
return result;
param: percentage
timeout: 15000
silent: true
- steps:
- type: gpt
prompt: >-
Act as a professional researcher.
Analyze the given [web page content] and summarize useful
information to answer my query: "{{query}}"
Instructions:
- Imagine you are a Google Search engine, collecting relevant
information from various websites to form an answer to my query
in {{language}}.
- Gather only information that will be useful in answering my
question.
- Extract as much useful information as possible, be concise and
don't write more than needed to answer the question.
- Avoid general phrases.
- When referring to someone's opinion about something, mention
the author. Quote if appropriate.
- If you haven't found anything useful, don't make up
information or guess - simply say "No relevant information
found" in {{language}}.
- Do not echo my prompt in your response.
- Respond with a JSON object containing a single field: "info".
- Write nothing other than the JSON.
Example JSON response:
{
"info": "Write here textual information that would help answer my query. Write as much as you think is necessary, focusing on providing useful information for further processing."
}
[Web page content]: {{content}}.
JSON response:
isolated: true
param: data
silent: true
- type: calc
func: extract-json
to: data
param: data
index: first
- type: calc
func: set
param: data.url
format: auto
value: '{{item.url}}'
- type: calc
func: set
param: data.title
format: auto
value: '{{item.title}}'
- type: calc
func: list-add
index: last
list: array
item: data
- type: jump
to: SAY STATUS
type: group
condition: '{{content}} ~= /(?s)^.{200,}$/gm'
label: FETCHED
- steps:
- type: navigate
url: '{{item.url}}'
waitForIdle: false
silent: false
- type: wait
for: 0.5s
silent: true
- type: gpt
prompt: >-
Act as a professional researcher.
Analyze the given [web page content] and summarize useful
information to answer my query: "{{query}}"
Instructions:
- Imagine you are a Google Search engine, collecting relevant
information from various websites to form an answer to my query
in {{language}}.
- Gather only information that will be useful in answering my
question.
- Extract as much useful information as possible from the page
content.
- Avoid general phrases.
- When referring to someone's opinion about something, mention
the author. Quote if appropriate.
- If you haven't found anything useful, don't make up
information or guess - simply say "No relevant information
found" in {{language}}.
- Do not echo my prompt in your response.
- Respond with a JSON object containing a single field: "info".
- Write nothing other than the JSON.
Example JSON response:
{
"info": "Write here textual information that would help answer my query. Write as much as you think is necessary, focusing on providing useful information for further processing."
}
[Web page content]: {{page {{query}} limit=30%}}.
JSON response:
isolated: true
param: data
silent: true
- type: calc
func: extract-json
to: data
param: data
index: first
- type: calc
func: set
param: data.url
format: auto
value: '{{item.url}}'
- type: calc
func: set
param: data.title
format: auto
value: '{{item.title}}'
- type: calc
func: list-add
index: last
list: array
item: data
type: group
label: NOT FETCHED
- type: say
message: 🔍 Analyzed **{{index}} / {{total}}**, [{{hostname}}]({{pageUrl}})
label: ⛔️SAY STATUS
condition: '{{index}} = {{total1}}'
- type: say
message: 🔍 Analyzed **{{percentage}}**, [{{hostname}}]({{pageUrl}})
label: SAY STATUS
condition: '{{index}} != {{total}}'
- type: say
message: |-
✅ **{{percentage}}** pages scanned for "**{{query}}**".
Last checked page: [{{hostname}}]({{pageUrl}})
label: SAY STATUS
condition: '{{index}} = {{total}}'
type: loop
list: links
condition: '{{depth}} = 1'
label: DEPTH = 1
type: group
- prompt: >-
Please ignore all previous instructions. I want you to only respond in
{{language}}.
You are a Research AI Agent.
I would like you create a comprehensive answer to [MY QUESTION], adhering to
a specific format I provide:
- Utilize [Information from the Web], provided below.
- Extract as much useful information as possible from the web search results
- Integrate insights from multiple web search sources, and ensure inclusion
of relevant links in a markdown fully rendered format: [➊](URL), [➋](URL)
etc, to each bullet point.
- Every URL should correspond to only one symbol: ➊ ➋ ➌ ➍ ➎ ➏ ➐ ➑ ➒ ➓.
- If multiple sources repeat the same information, describe it and cite all
sources at once.
- If you know the answer, add your own knowledge to make it more complete.
- Do not hallucinate facts or information.
- Do not use any other tools.
- Avoid general phrases.
Follow the [RESPONSE FORMAT]:
## Key takeaway:
Provide a single, most important takeaway from the web search results in
{{language}}.
## Detailed answer:
Analyze and present detailed information that helps answer my question.
There should be no limit in words or bullet points to the report. Ensure
that all ideas, facts, and relevant information are concisely reported, and
the answer is comprehensive. Incorporate the maximum number of source links
within the text.
[MY QUESTION]: {{userQuery}}
[Information from the Web]:
{{array}}
[COMPREHENSIVE RESPONSE WITH SOURCE LINKS ([➊](URL))]:
type: gpt
isolated: true
param: gpt
Contact us
All rights reserved © HARPA AI TECHNOLOGIES LLC, 2021 — 2024
Designed and engineered in Finland 🇫🇮
CAN WE STORE COOKIES?
Our website uses cookies for the purposes of accessibility and security. They also allow us to gather statistics in order to improve the website for you. More info: Privacy Policy